OpenAI today released a new generation of GPT-5.6 series models, including the flagship model Sol, the balanced model Terra for daily office use, and Luna, which focuses on speed and cost advantages, and is open to testing in a "limited preview" form for some partners. Among them, Terra is close to GPT-5.5 in overall performance, but the price is reduced to about half of the original price; Luna uses "strong capabilities" to match the current lowest price, targeting enterprises and developers who are highly cost-sensitive but still need medium and high-level large model capabilities.
As the focus of this update, GPT-5.6 Sol is positioned by OpenAI as the most powerful model to date, with significant enhancements in multiple "agent" capabilities such as coding, life sciences, and network security. Sol has added a "max" level inference intensity option and introduced an "ultra" mode, which can handle more complex task scenarios through internal sub-agent collaboration, targeting automated workflows in high-risk industries and complex system environments.
In terms of security, OpenAI emphasizes that GPT-5.6 Sol is equipped with the most "robust" security protection system to date, focusing on protecting high-risk activities, sensitive network security requests, and potential abuse scenarios. OpenAI said the model has been intensively tested against real-life attack scenarios, with the goal of providing substantial assistance to defenders, compliance security teams, and developers while limiting its potential for offensive use. According to the official introduction, GPT‑5.6 Sol is better at helping users discover and repair system vulnerabilities than performing end-to-end attacks.
It is worth noting that this release of the GPT‑5.6 series was subject to administrative intervention by the Trump administration in the United States and is currently not fully open to all users. According to OpenAI, the company has agreed to hold off on releasing full access to GPT‑5.6 to the general public at this stage, and will only provide access in the form of APIs and Codex to a small group of "trusted partners" and institutions. OpenAI stated that it will not consider the wider introduction of the GPT‑5.6 series of models in ChatGPT, Codex, and public APIs until more testing is completed, and the timetable is summarized as "coming soon."
In an official statement, OpenAI rarely publicly expressed reservations about the current government review process, believing that this type of government-led access mechanism should not become the long-term norm for future AI model releases. OpenAI points out that this approach will long-term state-of-the-art tools away from the users, developers, enterprises and network defenders who actually need them. However, the company also emphasized that the reason for accepting this arrangement in the short term is out of practical considerations in exchange for a broader open path in the future, and at the same time, it is in line with the assessment framework and reusable review process established by the Trump administration around the cybersecurity executive order.
In accordance with the latest executive order signed by the Trump administration on June 2, regulatory authorities will establish a pre-launch benchmarking and evaluation mechanism for new generation AI models, with a particular focus on cybersecurity and high-risk application areas. Prior to this, Anthropic, an AI company that has also been subject to US regulatory attention, has been required to remove Claude Fable 5 and Mythos 5-related access capabilities, although the company has voluntarily submitted to government review and added new safety "guardrails" based on feedback. This also reflects that the current US regulatory attitude towards the release of cutting-edge large models is becoming more cautious and more interventionist.
Currently, the GPT‑5.6 series models are only open to a group of partners and institutions selected by OpenAI through API and Codex, and ordinary users cannot experience them directly in ChatGPT for the time being. OpenAI stated that after further consultation and testing processes with the government are completed, it will promote wider opening in the next few weeks and gradually integrate the GPT‑5.6 series into the existing product matrix.